Research map / Label-efficient and robust learning / Semi-supervised learning
Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AmbiSSL, a semi-supervised medical image segmentation framework that uses a few multi-annotator labels plus unlabeled scans to produce diverse plausible segmentations via randomly pruned decoders, a shared latent space, and cross-decoder supervision.
Problem
Medical image segmentation needs costly annotations, and models usually output a single deterministic mask even though expert annotations can disagree; existing ambiguous segmentation methods require multiple annotations per image and cannot use unlabeled data.
Approach
Creates diverse pseudo-label sets from multiple decoders obtained by randomly pruning the final decoder layers of a backbone, learns a common latent space from labeled multi-annotator data and unlabeled pseudo-labels (Gaussian posterior for labeled, Laplace for unlabeled), and applies cross-decoder supervision so pruned decoders guide each other.
Key results
- On LIDC-IDRI with 10% labeled data, AmbiSSL reaches GED 0.1620 and soft Dice 89.86%, beating the best baseline (Baseline III, GED 0.1779, soft Dice 88.75%).
- On ISIC with 20% labeled data, AmbiSSL reaches soft Dice 85.87% and GED 0.2444, versus 83.32% and 0.2732 for the best baseline (Baseline II).
- Ablation shows alpha_u = 0.5 performs best; larger alpha_u = 1.0 degrades GED and soft Dice on both datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AmbiSSL instead of Baseline III | LIDC-IDRI semi-supervised segmentation with 5% labeled data · LIDC-IDRI, 5% labeled | Dicesoft: 87.29 -> 88.34 | Helped |
| use AmbiSSL instead of Baseline III | LIDC-IDRI semi-supervised segmentation with 10% labeled data · LIDC-IDRI, 10% labeled | GED: 0.1779 -> 0.1620 | Helped |
| use AmbiSSL instead of Baseline II | ISIC semi-supervised segmentation with 10% labeled data · ISIC 2018, 10% labeled | Dicesoft: 81.54 -> 83.40 | Helped |
| use AmbiSSL instead of Baseline II | ISIC semi-supervised segmentation with 20% labeled data · ISIC 2018, 20% labeled | GED: 0.2732 -> 0.2444 | Helped |
| set αu to 0.3 instead of 0.5 (ablation) | AmbiSSL · LIDC-IDRI, 5% labeled | GED: 0.1872 -> 0.1987 | Hurt |
| set αu to 1.0 instead of 0.5 (ablation) | AmbiSSL · LIDC-IDRI, 5% labeled | GED: 0.1872 -> 0.2134 | Hurt |
| set αu to 0.3 instead of 0.5 (ablation) | AmbiSSL · LIDC-IDRI, 10% labeled | Dicesoft: 89.86 -> 88.95 | No clear effect |
| set αu to 1.0 instead of 0.5 (ablation) | AmbiSSL · LIDC-IDRI, 10% labeled | Dicesoft: 89.86 -> 85.88 | Hurt |
| set αu to 0.3 instead of 0.5 (ablation) | AmbiSSL · ISIC 2018, 10% labeled | Dicesoft: 83.40 -> 83.52 | No clear effect |
| set αu to 1.0 instead of 0.5 (ablation) | AmbiSSL · ISIC 2018, 10% labeled | Dicesoft: 83.40 -> 79.65 | Hurt |
| set αu to 0.3 instead of 0.5 (ablation) | AmbiSSL · ISIC 2018, 20% labeled | Dicesoft: 85.87 -> 85.32 | No clear effect |
| set αu to 1.0 instead of 0.5 (ablation) | AmbiSSL · ISIC 2018, 20% labeled | Dicesoft: 85.87 -> 81.19 | Hurt |
Surprising findings
- Prior fully-supervised ambiguous methods such as Pionono and Probabilistic U-Net perform much worse on ISIC at low label ratios than semi-supervised baselines.
- Pionono reaches a soft Dice of only 52.15% on ISIC with 10% labels, far below the other ambiguous methods.
Limitations
- Performance remains below fully annotated methods, as the authors acknowledge.
- Evaluated only on two public datasets (LIDC-IDRI and ISIC 2018) with 2D images.
- Results on ISIC with 5% labels were unstable so only 10% and 20% were reported.
Experimental setup
Datasets: LIDC-IDRI ISIC 2018
Benchmarks: LIDC-IDRI ISIC 2018
Metrics: Generalized Energy Distance soft Dice Dice max Dice match
Baselines: Probabilistic U-Net CM-Global CM-Pixel Pionono D-Persona Self-training (Baseline I) Cross-pseudo supervision (Baseline II) Consistency regularization (Baseline III)
Builds on: Probabilistic U-Net Cross pseudo supervision Pionono Generalized Energy Distance
Closest papers by meaning
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.88)
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.87)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.87)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.87)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.86)
- Diversified and Personalized Multi-rater Medical Image Segmentation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.